The narrative that the cloud market has sold for years has been one of commoditization: the hardware underneath doesn't matter, what matters is the service. You chose the provider based on the catalog of managed services, the price per instance hour, and the support. The chip was an infrastructure detail, invisible and interchangeable — convenient while everyone was buying the same silicon from Intel and AMD and the real difference was in software and operations. The problem is that this narrative is no longer true, and providers who understood this before are reaping benefits that others cannot afford.
Why build your own chip when you can buy it
The direct answer: because at hyperscaler scale, hardware efficiency turns into margin numbers that justify any reasonable R&D investment. AWS processes compute volumes that make any percentage gain per watt relevant to hundreds of millions of dollars per year. If a chip designed specifically for your workloads delivers twenty percent more performance per watt than the off-the-shelf equivalent, that difference isn't a technical curiosity — it's a competitive advantage converted into price, margin, or both.
The second reason is independence. Anyone who buys all computing capacity from a single supplier — in the case of AI, essentially from Nvidia — is subject to the allocation queue, the price dictated by whoever has a practical monopoly on the product, and the technological roadmap of another. For a company competing in cloud computing, this is a strategic risk in the same category as dependence on a single customer. No one on the council would tolerate the second; the first became routine until big techs decided on the side of their own production.
How each hyperscaler played this hand
AWS Graviton is the most readable case as a business decision. Based on ARM, Graviton4 delivers significant performance gains per dollar over equivalent x86 instances on AWS itself — and AWS uses this both to reduce internal costs and to offer cheaper instances to migrating customers. Adoption has grown because the value proposition is quantifiable: you run the same workload and pay less, or you run more for the same price.
Google made the most radical move with TPUs. The Tensor Processing Unit was designed from scratch for a specific workload: high-dimensional matrix multiplication, which is the core operation of neural networks. TPU v5p is used for both training and inference, deeply integrated with TensorFlow and JAX. The relevant detail for anyone considering using it: the TPU does not exist as isolated hardware. You access capacity through Google Cloud — adopting TPU is a platform decision, not just a compute decision. Performance is linked to the bond.
Apple is the most extreme case of vertical integration and the oldest. Since the M1, Apple has controlled the chip, operating system, compiler, and system software. The practical consequence is that optimizations that would be impossible in a fragmented stack become routine — the chip knows about the cache, the operating system knows about the chip, the compiler knows about both. The M4 Ultra achieves server performance in a desktop form factor not by a single engineering miracle, but because each layer is designed for the others.
Meta and Microsoft arrived later with more surgical goals. Meta's MTIA was designed to reduce reliance on Nvidia specifically for recommendation models — billions of inferences per day at a standard stable enough to justify specialized silicon. Microsoft's Maia 100 covers training and inference on Azure, including some of the GitHub infrastructure Copilot. In both cases, the goal is not to replace GPU entirely, but to cut dependency where predictable load makes specialization economically obvious.
The chip as moat: why this is hard to copy
What makes custom silicon a durable advantage is the cost and time to replicate. Designing a competitive chip takes years before the first wafer leaves the factory. Manufacturing at scale requires relationships with TSMC or Samsung that are not built in quarters. The software ecosystem that gets the most out of the chip takes even longer. Whoever started in 2016 — Google with the TPU, Apple with its transition chips — is in a position that a competitor who decides to start tomorrow won't reach before 2030 in the best case scenario. It is not an advantage that can be bought in an investment round or copied in a product sprint.
For smaller cloud providers, this creates a competitive problem with no obvious solution. You can buy the same Nvidia or AMD chips as the competitor, but you can't buy the efficiency that Graviton gives AWS after years of full-scale iteration. The performance-per-dollar gap widens with each generation — and there is no shortcut.
What does this change for those who use the cloud, not for those who sell it
The most relevant change for those who consume cloud services is that the instance performance benchmark has lost a dimension. Comparing vCPUs and gigabytes of RAM between providers as if the underlying hardware were equivalent is an incomplete analysis. AWS Graviton instances and equivalent AWS x86 instances are different workloads on different silicon, with performance standards that vary by application type.
The second change is that cloud lock-in gained a hardware layer that was not in the original contract. An application optimized for TPU has a much higher migration cost than one on generic compute. This is not an argument for avoiding specialized chips — performance often justifies it — but it is cost that needs to be included in the analysis before adoption, not after.
The third dimension is that differences in cost and performance between providers will grow. Graviton already makes some loads materially cheaper on AWS than on off-the-shelf hardware alternatives. As each hyperscaler matures its silicon, comparing clouds will require understanding which chip is behind which workload — technical knowledge that many companies today delegate to the integrator on call, without realizing that they are also delegating the decision to the provider.
Also read
- Inference chips and ASIC: when the specialized beats the generic
- Specialized chips and the end of the generic CPU era
- Edge computing: why computing is leaving the cloud and moving closer to the data
- Silicon geopolitics: why the chip war matters for those who decide about technology
- Inference at the edge: when running AI on the device makes more sense than in the cloud
- Infrastructure as a competitive advantage: what startups learn from big techs